PhD Student — University of Alberta

Devon
Yanitski

I study minds and language models: psychometrics from language embeddings, the neural correlates of human memory, and LLM interpretability tools applied to human minds.

Department of Psychology, University of AlbertaComputational Memory Lab

Psychometrics from embeddings

Can the geometry of an embedding space stand in for human survey responses? Semantic Factor Analysis recovers latent psychological structure from item text alone.

Tool

semanticfa

Devon Yanitski, Chris Westbury

An R package for response-free semantic analysis of psychometric scales — a full toolkit that recovers and refines a scale's latent structure from item wording alone. Factor analysis on item embeddings is the centerpiece, alongside semantic similarity matrices, factor retention, semantic 'loadings', theory comparison, redundancy / jingle-jangle checks, and short-form building. No human response data required.

Rfactor analysisembeddingspsychometrics

CRAN

Big Five · embedding space

drag to rotate

loading embedding space…

50 Big-Five items embedded by Qwen3-Embedding-8B (4096-d), projected to 3D with UMAP on cosine geometry. Items from the same factor cluster together, showing that the factor structure can be recovered from meaning alone.

What's latent in language models

Opening up language models — sparse-autoencoder feature steering, and what clamping a single interpretable feature reveals about the computations running inside.

loading the rat…

Qwen3.5-2B with a slider to steer the sparse-autoencoder feature corresponding to rats (#26631). A rat-themed take on Golden Gate Claude. This is a small LLM and may produce repetitive output.

sparse autoencodersfeature steeringQwen3.5-2BRats

GitHub →

Paperunpublished

A Computational Analog of OCD in LLMs Through Sparse-Autoencoder Feature Steering

Devon Yanitski

OCD is defined by intrusive, unwanted thoughts a person cannot dismiss on command and that grow stronger when suppressed. We use sparse-autoencoders (SAEs) to amplify features in LLMs that represent obsessive concepts. This produced the defining marks of clinical obsession: the model could not suppress the theme when instructed to, whereas the same theme from ordinary prompting stopped on command, and telling the model not to think about the theme raised the feature's activity further.

computational psychiatrySAE steeringOCDinterpretability

Are LLMs dual-process thinkers?

Do language models choose like utilitarians, and does reasoning shift their judgements the way it shifts ours? Measuring when LLMs replicate human behaviour.

Paperunpublished

Reasoning in LLMs Causes More Utilitarian Judgements

Devon Yanitski

Does the human dual-process finding — that reasoning increases utilitarian moral judgment — extend to LLMs? Across 122 open-source models run on a 40-item moral-dilemma battery, switching on reasoning raised utilitarian responding and lowered altruism and deontology: the human dual-process signature, reproduced in a system with no affect, kin, or evolved morality.

moral psychologyreasoning122 LLMsprocess dissociation

Paperunpublished

LLMs Validate The Cognitive Reflection Test

Devon Yanitski

122 open-source LLMs took an updated Cognitive Reflection Test (Meyer et al., 2024), each item answered twice — once on intuition, once with reasoning. Holding arithmetic ability fixed, engaging reasoning alone moves the score, supporting the claim that the CRT is separable from ability — a cleaner demonstration than human data, where CRT scores confound ability, disposition, and motivation.

cognitive reflection testreasoning122 LLMsMeyer 2024

Improving LLM-assisted research

I am building the scaffolding that lets AI agents carry out autonomous psychology research — starting with how they ground themselves in the literature.

Tool

AP-RAG

A major roadblock preventing AI agents from performing autonomous research are hallucinations — when an LLM lacks the full context of an area of research, it often makes a plausible yet incorrect guess. However, it isn’t possible to dump all relevant prior research into an LLM's limited context window. An academic-paper retrieval-augmented generation (AP-RAG) system solves both problems: it lets AI agents efficiently retrieve the relevant chunks of information from a large body of research.

graph RAGacademic papersretrieval

GitHub

AP-RAG Papers Atlas

atlas

The semantic data of the 10,000+ papers in my personal database. Papers with similar meanings are grouped together; higher peaks means more papers. Agentic LLMs can navigate this space to understand my area of research.

drag to orbit · scroll to zoom · esc to reset

Viewing the live brain

Real-time EEG tooling — Muse Visualizer turns the raw waveforms and oscillatory band activity from a consumer headband into highly customizable, artistic live representations of the brain.

loading clip…

Playing the Council of Elrond theme (LOTR) on piano while my live brain activity renders in Muse Visualizer.

open ↗

Tool

Muse Visualizer

Ultra-high-performance (200+ FPS) real-time EEG visualization for the Muse headband — raw waveforms and Delta/Theta/Alpha/Beta/Gamma band power over Lab Streaming Layer.

PyQtGraphLSLsignal processingPPG

GitHub

Tools & toys

Tool

Slurm HPC Dashboard

A VS Code extension showing real-time Slurm job status and fairshare trends across HPC clusters in the status bar and a rich dashboard.

TypeScriptVS Code ExtensionSLURMHPC

GitHub

Game

Txtr

A Temple-Run-inspired typing game about texting and driving — dodge five lanes of traffic with the arrow keys while typing scripted replies. Vanilla HTML/CSS/JS on a canvas.

JavaScriptcanvastyping game

GitHub

Tool